codex-1 introduced RLEF (Reinforcement Learning from Execution Feedback), where code generated by the model is actually executed and the test results are fed back to the model as reward signals.